CAREER: Scalable Algorithms for Extreme Computing on Heterogeneous Hardware, with Applications in Fluids and Biology

目前，全世界都在寻求在本世纪末实现亿级计算，美国、中国和日本尤其如此。在美国，总统的美国创新战略(2009)明确列出了通过艾级计算机大幅提高我们的模拟能力的目标。实现这一目标的挑战是前所未有的：功率限制，微芯片制造达到物理极限，计算能力和互连带宽之间日益失衡，以及系统中核心数量的不断增加。当务之急是开发可扩展的算法，以利用新系统的巨大并行性，并培养下一代计算科学家。其中第一个是这个职业项目的科学部分的中心。一种潜在的变革性组合正在出现，其中一类分层算法提供了与问题大小成线性关系的理想伸缩性，并以优异的性能映射到多核硬件(如GPU)。将对算法进行改进，例如合并树形编码和快速多极子方法的元素、避免通信和同步、动态误差控制和自动调整计算。该研究计划是跨学科的垂直整合，包括在流体动力学和生物系统的极端规模的应用。该项目将产生高度可扩展的科学软件，重新制定算法，以在多核硬件上实现最大性能。通过开源模式传播和管理，提供的计算基础设施将提供最大影响，超出重点应用领域。能够扩展到数百万处理器的社区软件将是开发后千万亿级系统的关键，本项目旨在提供这一点。另一方面，这一计划的教育部分建立在PI在使用技术支持学习以及在催化国际合作和推广方面的成功记录的基础上。该计划包括改善教育环境，利用技术进行课程教学，并(通过开放课程软件)促进国家的科学素养。将通过校外高级培训活动和在线学习媒体实现培养下一代计算科学家的目标。

There is currently a world-wide quest to achieve exascale computing by the end of the decade, with vigorous efforts in the US as well as China and Japan, in particular. In the US, the President's Strategy for American Innovation (2009) explicitly lists among its goals to dramatically increase our simulations capacity via an exascale computer. The challenges to achieve this goal are unprecedented: power constraints, microchip fabrication reaching physical limits, the growing imbalance between compute capacity and interconnect bandwidth, and the ever increasing number of cores in a system.Among the matters of highest priority are development of scalable algorithms that can exploit the enormous parallelism of new systems, and educating the next generation of computational scientists. The first of these is at the center of the scientific part of this CAREER project. A potentially transformative combination is emerging where a class of hierarchical algorithms, offering ideal scaling linear with problem size, maps with excellent performance to many-core hardware (such as GPUs). Algorithmic improvements will be undertaken, such as combining elements of treecodes and fast multipole methods, communication and synchronization avoidance, dynamic error control and auto-tuning of the computation. The research program is vertically integrated across disciplines, including applications at extreme scales in fluid dynamics and biological systems.This project will produce highly scalable scientific software, reformulating the algorithms to achieve maximum performance in many-core hardware. Disseminated and curated via the open-source model, the computational infrastructure delivered will offer maximum impact, beyond the application areas of focus. Community software that is able to scale to millions of processors will be crucial to exploit post-petascale systems, and this project aims to provide that. The educational part of this program, on the other hand, builds on the PI's track record of success both in the use of technology to support learning, and in catalyzing international collaboration and outreach. The program includes enhancing educational environments using technology for both curricular instruction and contributing to the nation's science literacy (via open courseware). The goals of fostering the next generation of computational scientists will be pursued via extra-mural advanced training events, and online learning media.